Few-shot Face Recognition Based On Deep Transfer Learning

Despite recent advances in hot domains such as computer vision,the conventional supervised deep learning networks are not applicable for scarce data.The prototypical deep learning architectures in the field of face recognition depend much on large datasets and repetitive training.Therefore,this paper makes extensive studies on few-shot face recognition based on the latest development of deep learning and deep transfer learning and obtain great results.To begin with,we summarize some conventional and newest theories about face recognition.And we analyze the merits and demerits of nowadays methods to draw forth the few-shot problem we are facing with.To tackle this issue,we employ ideas from transfer learning based on deep neural features.Our framework learns a few-shot small neural network that maps a small labelled training set to large unlabeled test set.Once the network has been constructed,we can acquire strong face feature representation on pretrained network and discriminate subtle face features to realize category generalization.We define this few-shot algorithm problem on some public datasets and on the dataset established by our own which achieves strong results compared to competing approaches.Subsequently,we explore the transferrable ability of the pretrained model based on modified few-shot small neural network and choose the best architecture based on layer transfer.Experiments show that combined few-shot samples with auxiliary datasets,the optimized transferring model is superior to other deep learning-based methods which achieves good results.Finally,we design a holistic end-to-end system integrating face detection and alignment,clustering and recognition,which handles videos and images as input in a timely manner.Secondly,we add face clustering in the following part of face detection.And the clustering step is conducted to find face relations based on Interpretive Structure Model to divide faces into groups,avoiding duplicate feature computation and repetitive face recognition.And we establish database to save all useful information laying foundation for data retrieval and intelligence analysis.Experiments show that the system can efficiently process video and images,and accurately recognize few-shot faces.